A multi-view and retrieval-enhanced graph-based influence-conducting analysis pattern discovery method

Abstract

This invention relates to a method for discovering influence transmission patterns based on multi-perspective and retrieval-enhanced graphs. This method combines multi-perspective influence argument graph analysis with retrieval-enhanced generation (RAG) technology. Key steps include multi-level abstraction analysis of the influence argument graph from fine-grained to coarse-grained, multi-dimensional abstraction analysis of RAG enhancement, and direct and indirect influence path analysis with variable step sizes in RAG enhancement. This enhances the contextual understanding of the influence argument graph, improves the comprehensiveness of influence analysis, and increases the accuracy and depth of influence transmission pattern discovery. In this method, the RAG system is responsible for retrieving specific contextual evidence, assisting in classifying influence argument graph elements, understanding abstract dimensional influences, and verifying influence paths. Furthermore, with the enhancement of retrieved evidence, the generative large language model (LLM) can comprehensively generate a comprehensive and accurate influence transmission model of major events, achieving efficient discovery of complex influence transmission patterns.

Description

Technical Field

[0001] This invention relates to the fields of automated data analysis and knowledge discovery. More specifically, this invention relates to a method for discovering patterns based on influence transmission analysis using multi-view and retrieval-enhanced graphs. Background Technology

[0002] Tracking the propagation of impacts within interconnected events or networks of entities is a complex problem. The transmission paths of impacts are often not direct or obvious, making the identification of indirect or implicit impact paths particularly difficult. For example, in emergency response, geopolitics, or the economic sphere, an initial event can, through a series of intermediate links, ultimately have a profound impact on seemingly unrelated entities or systems. Furthermore, abstracting and generalizing universal patterns of impact transmission from specific historical event data is crucial for predicting future event evolution and developing response strategies, but this itself is fraught with challenges.

[0003] Graph structures provide a powerful framework for representing complex relationships and influence transmission between entities. However, traditional graph-based analysis methods have some inherent limitations. Many existing graph traversal algorithms, such as breadth-first search (BFS) or depth-first search (DFS), while capable of discovering paths, often lack an understanding of the deep semantics or contextual awareness behind nodes and relationships. Although these algorithms can identify connections, they cannot explain the nature or strength of these connections.

[0004] Furthermore, static graph analysis struggles to capture the dynamics or evolutionary characteristics of influence patterns. Influence transmission is often a time-varying process, and static graph models may fail to adequately reflect this dynamism. Simultaneously, effectively integrating diverse, unstructured information sources (such as news reports, policy documents, and social media comments) into graph analysis to provide a more comprehensive context is also a challenge faced by existing methods.

[0005] Retrieval Augmented Generation (RAG) technology significantly enhances the ability of Large Language Models (LLMs) to handle knowledge-intensive tasks by combining them with external knowledge bases, thereby improving the accuracy and reliability of responses. RAG technology provides additional knowledge input to LLMs by accurately retrieving and integrating document fragments highly relevant to the user's query, enabling them to generate more factually supported and context-rich responses.

[0006] However, traditional RAG applications may not be optimal for sophisticated tasks such as complex graph analysis and the discovery of influence transmission patterns. For example, "Naive RAG" may fail to extract sufficient information when dealing with graph data with intricate and interconnected influence relationships due to its imprecise retrieval. Nodes and edges in graph data inherently contain rich structural and semantic information, which simple text block retrieval may not fully utilize. More importantly, general-purpose RAGs lack targeted design for specific analytical perspectives, making it difficult to fully understand or utilize the implicit analytical perspectives in documents, such as influence analysis with multiple levels of abstraction, multiple dimensions of abstraction, and variable path step sizes. These are crucial for in-depth discovery of influence transmission patterns. Simply applying general-purpose RAGs to graph data without adjusting and optimizing for these specific analytical perspectives will make it difficult to reveal deep-seated influence transmission mechanisms. Therefore, a more specialized RAG integration method is needed, which possesses stronger capabilities for understanding complex graph data and enhances the multi-perspective graph analysis framework to discover the deep-seated influence transmission patterns hidden within graph data.

[0007] In conclusion, there is an urgent need for an advanced research method that can synergistically combine the rigor of structured graph analysis with the contextual understanding and generative capabilities of RAG. This method should overcome the limitations of existing technologies, not only identifying influence transmission paths but also gaining a deeper understanding of the underlying patterns, mechanisms, and multiple factors, thereby providing a more powerful and accurate tool for influence transmission analysis in complex systems. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide a method for discovering influence transmission patterns based on multi-view and retrieval-enhanced graphs. By combining multi-view influence argument graph analysis with retrieval-enhanced generation (RAG) technology, it can achieve influence pattern analysis with multiple abstract levels, multiple abstract dimensions, and variable transmission path step sizes, thereby improving the ability to understand complex graph data and discover complex influence transmission patterns.

[0009] Technical Solution: To achieve the above objectives, the present invention provides a method for discovering influence transmission patterns based on multi-view and retrieval-enhanced graphs, comprising the following steps:

[0010] Step 1: Construct an impact argument graph based on event data. The nodes of the impact argument graph represent events or impacts, and the edges represent the transmission of impacts.

[0011] Step 2: Establish a RAG system, which includes a knowledge base storing structured and unstructured information related to the event data, a retrieval module, and a generative large language model;

[0012] Step 3: Based on the RAG system, perform RAG-enhanced multi-level abstraction analysis on the influence argument diagram;

[0013] Step 4: Based on the RAG system, perform RAG-enhanced multi-dimensional abstraction analysis on the influence argument diagram;

[0014] Step 5: Based on the RAG system, perform RAG-enhanced variable step size influence path analysis on the influence argument graph;

[0015] Step 6: Based on the RAG system, perform RAG-contextualized influence strength calculation on the influence argument graph;

[0016] Step 7: Based on the impact patterns and impact paths identified in Steps 3-6, generate a comprehensive model of the transmission of the impact of major events.

[0017] Preferably, in step 3, multiple abstraction levels are defined, ranging from fine-grained to coarse-grained generalization levels. RAG-enhanced multi-level abstraction analysis is performed on the impact argument graph at different granularities to enhance the contextual understanding of major emergency events and their impacts in the impact argument graph.

[0018] Preferably, in step 3, for the abstract level The retrieval module of the RAG system is used to retrieve specific contextual information of the argument graph, including: targeting nodes in the argument graph. Retrieval and Abstraction Level Feature matching and supports nodes Divided into abstract levels Specific contextual information; among which, the introduction of RAG makes the analysis process more dynamic, accurate, and insightful; for the level of abstraction Search criteria The input is the nodes in the argument graph. and nodes Corresponding event description The output is a boolean value, representing the node. Does it belong to the abstract level? The formula for judging the event is:

[0019] .

[0020] Preferably, in step 3, for the abstract level Using the generative large language model of the RAG system The specific contextual information is used to identify and generate the representation of the abstraction level. The influence transmission pattern of characteristics; once an event node is categorized into a specific level of abstraction, the generative large language model analyzes the level of abstraction. The impact patterns are analyzed and descriptions are generated; the analysis at each level of abstraction includes RAG query examples, retriever behavior, and generator behavior.

[0021] Preferably, in step 4, the RAG system is used to identify multiple predefined abstract dimension attributes that influence the nodes of the argument graph, analyzing the influence of the argument graph from multiple abstract dimensions to improve the comprehensiveness of the analysis; the abstract dimensions include at least time and location dimensions, and each abstract dimension contains multiple predefined categories, for example, the time dimension includes short-term ( ), medium term ( ),long( The location dimension includes coastal areas ( ), mountainous regions ( ), plains area ( ), urban areas ( For a specific abstract dimension category to which an event node belongs, the retrieval module of the RAG system is used to retrieve contextual information related to that specific abstract dimension category, including the general influence mechanism or historical cases of the influence transmission of that dimension category.

[0022] Preferably, in step 4, the generative large language model is used in conjunction with the contextual information related to the specific abstract dimension category to analyze and explain how the specific abstract dimension affects the event transmission pattern. The analysis for multiple abstract dimension categories includes RAG query examples, retrieval behavior, and generator behavior.

[0023] Preferably, step 5 defines at least a long step size. and short step length Step size for analysis of multiple influence paths ( The system performs direct and indirect impact path analysis with variable step size. Using the RAG system, it retrieves evidence of causal relationships supporting the connections between nodes in the impact path. The long step size focuses on the direct and macroscopic impact transmission relationship between the starting node and the final node, ignoring intermediate nodes. The short step size focuses on the asymptotic transmission path that includes intermediate nodes, considering the connection from the starting node to the final node through some intermediate nodes.

[0024] Preferably, in step 5, for a given starting and ending node, the RAG system is used to assist in the discovery and verification of influence paths. Candidate paths are identified in the argument graph using a graph search algorithm, and the retrieval module of the RAG system is used to search the knowledge base for textual evidence supporting the candidate paths, discovering and verifying that they conform to the analysis step size of each influence path. Defined potential impact paths;

[0025] The influence path discovery and verification assistance includes RAG query examples, retriever behavior, and generator behavior, for step size... Search criteria The input is a debate diagram. Starting node and end node The output is the combination of path nodes. and the influence transmission relationship The calculation expression is:

[0026]

[0027] Preferably, in step 5, the generative large language model is used in conjunction with the search for textual evidence supporting candidate paths from the knowledge base to evaluate the transmission mechanism of path influence; the evidence enhancement of the transmission mechanism of path influence includes RAG query examples, retrieval behavior, and generator behavior.

[0028] Preferably, in step 6, the RAG system is used to calculate the intensity of contextualized influence for combinations of influencing path nodes. Node sequence in Influence on transmission relationship Calculate the influence strength between nodes To indicate, the The calculation formula is as follows:

[0029]

[0030] in, and It is a node and The components, represented by a certain semantic vector, are converted into semantic vectors through natural language processing to transform node information. It is a logical relation weight, determined based on the causal relationship and logical connection between events. For example, if event Directly caused the incident , Take the larger value, and vice versa.

[0031] In step 6, utilizing the aforementioned RAG contextualized impact strength, the representation of the semantic vector for each node (event / impact) is dynamically enriched or adjusted based on the current context. In implementation, this is done at the computation node... and Before assessing the strength of the influence between them, information about the RAG system can be queried. and Specific information within the context of the currently analyzed relationship. For example, if in the analysis right For impacts within a specific time period or geographical environment, RAG should retrieve specific details relevant to this scenario. The retrieved information can be used to generate richer, more context-aware datasets. and The temporary embedding vector is specifically used to calculate the influence strength of this contextualization; or the pre-calculated general node embedding vector is adjusted or weighted so that the numerator of the formula for calculating cosine similarity more accurately reflects the context. and The degree of semantic relevance in the current specific context.

[0032] In step 6, the influence strength of the RAG contextualization described above is used as the logical relation weight. The assignment provides strong evidence to support this. Based on the causal relationships and logical connections between events, LLM can analyze textual evidence retrieved by RAG, thereby providing... To assign a more accurate and reasonable value.

[0033] Beneficial effects: The present invention has the following advantages:

[0034] 1. Improve the accuracy and depth of identifying influence patterns: Through the contextual information provided by RAG, the analysis is no longer limited to the structure of graph data and the nodes themselves, but can also understand the semantic meaning of nodes and relationships;

[0035] 2. Enhanced ability to discover weak or indirect effects: RAG can retrieve seemingly unrelated information from large amounts of data that actually contributes to the transmission of effects, helping to discover weak or indirect, insignificant effect paths;

[0036] 3. Enhanced contextual understanding: RAG's enhanced multi-perspective analysis enables a more comprehensive understanding of the influencing mechanisms, taking into account complex influencing factors with multiple levels and dimensions of abstraction, as well as the influence transmission patterns of different path step lengths, whether direct or multi-hop indirect.

[0037] 4. Generate more reliable and evidence-supported impact models: The impact transmission patterns and models generated by RAG-enhanced LLM are more persuasive because they are supported by factual evidence retrieved from the knowledge base;

[0038] 5. Adaptability to impact analysis in various fields: By designing a multi-perspective analysis framework, this method can be widely applied to various fields that require impact analysis, such as geopolitics, economics, and emergency response. Attached Figure Description

[0039] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0040] The technical solution of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0041] Example 1

[0042] like Figure 1 As shown in the figure, this embodiment presents a method for discovering influence transmission analysis patterns based on multi-view and retrieval enhancement graphs, which includes the following steps:

[0043] Step 1: Construct an impact argument graph based on event data. The nodes of the impact argument graph represent events or impacts, and the edges represent the transmission of impacts.

[0044] Step 2: Establish a Retrieval Enhanced Generation (RAG) system, which includes a knowledge base storing structured and unstructured information related to the event data, a retrieval module, and a generative large language model (LLM).

[0045] Step 3: Perform RAG-enhanced multi-level abstraction analysis on the impact argument graph. This invention utilizes the RAG system to dynamically retrieve and integrate specific contextual information, dividing events and impacts into different levels of abstraction from fine-grained to coarse-grained, to achieve a more accurate and detailed understanding of impact patterns at different granularities. The RAG system assists LLM in performing more precise hierarchical classification and pattern description by retrieving contextual evidence from the knowledge base that matches the characteristics of each abstraction level.

[0046] Step 4: Perform RAG-enhanced multi-abstract-dimensional analysis on the impact argument diagram. This invention utilizes the RAG system to acquire and integrate data related to abstract dimensions such as time and geography, such as historical cases or theoretical knowledge on how specific time windows or geographical environments affect the spread of events, providing LLM with more in-depth, richer, and context-aware abstract-dimensional impact analysis.

[0047] Step 5: Perform RAG-enhanced variable-step-size influence path analysis on the influence argument graph. This invention utilizes the RAG system to identify, retrieve, and evaluate supporting evidence for direct or multi-hop indirect influence paths of varying lengths in the argument graph. The RAG system can assist in discovering potential paths and extract information from a knowledge base to verify the validity of the paths, explain the transmission mechanisms, and assess their influence strength.

[0048] Step 6: Perform RAG-based contextualized influence strength calculation on the influence argument graph. This invention utilizes the RAG system to calculate the contextualized influence strength between nodes based on the current analytical context. This serves two purposes: firstly, to dynamically enrich or adjust the representation of node semantic vectors; and secondly, to provide evidentiary support for determining the weights of logical relationships.

[0049] Step 7: Collaborative RAG-argument graph integration for generating influence transmission patterns. Adjust the knowledge base, retrieval system, generator, and other components of the RAG system, focusing on graph structure analysis, to enable it to manipulate and utilize structured argument graphs, as well as the multi-perspective influence pattern and influence path analysis results from steps 3, 4, 5, and 6, to generate a more accurate, comprehensive, and context-based integrated model of the influence transmission of major events.

[0050] This invention combines Retrieval Enhancement Generation (RAG) technology with a multi-perspective graph analysis framework to achieve an understanding of complex graph data and the discovery of complex influence transmission patterns. Specifically, this method utilizes historical event data represented in the form of influence argument graphs. In the influence analysis stage, with multiple levels of abstraction, multiple dimensions of abstraction, and variable transmission path step sizes, RAG technology is used for retrieval enhancement, providing highly relevant and multi-perspective factual evidence to support graph data analysis and the description and induction of influence patterns.

[0051] Example 2

[0052] Based on the method provided in Embodiment 1, this embodiment provides a specific scenario: analyzing the impact of an event in which “country A in a certain region suffers a cyberattack on its port infrastructure” on international trade agreements and diplomatic relations.

[0053] Step 1: Construct an impact argument diagram based on event data.

[0054] (1) Identify key event E and construct nodes in the impact argument diagram. :

[0055] E1, Port X in Country A suffered a large-scale cyberattack, resulting in operational paralysis.

[0056] E2, Country B (which heavily relies on port X for trade) is severely disrupted in its import and export trade, and its supply chain is broken.

[0057] E3: The International Shipping Organization (ISO) issued a warning, raising the shipping risk level for the region.

[0058] E4. Country A publicly accused Country C of being the mastermind behind the cyberattack.

[0059] E5. Country C strongly denies Country A's accusations and counter-accuses Country A of diverting domestic problems.

[0060] E6: The G7 nations held an emergency meeting to discuss the impact of cybersecurity on global trade.

[0061] E7. Country B begins to seek a closer trade partnership with Country D and explores new logistics channels.

[0062] E8: An international forum proposed adding mandatory cybersecurity clauses to trade agreements.

[0063] E9. Diplomatic relations between country A and country C remain tense, and both countries recall their ambassadors.

[0064] (2) Constructing the edges of the influence argument graph:

[0065] Based on the logical sequence of events and the known preliminary causal relationships, edges are constructed to connect these nodes, forming an initial impact argument graph. For example, → , → , → , → wait.

[0066] Step 2: Establish a Retrieval Enhancement Generation (RAG) system, set up a RAG knowledge base, and store structured and unstructured information related to the event data.

[0067] (1) Data sources: collect and process news reports, official statements, think tank analysis reports, economic data (such as trade volume and port throughput) of relevant countries such as A, B, C, D, G7, International Shipping Organization Y, and relevant international forums, as well as cybersecurity technology analysis articles, historical cases of similar cyberattacks, existing international trade agreement texts, and diplomatic relations history.

[0068] (2) Processing: Divide the data into blocks, vectorize them, and build vector indexes and graph indexes.

[0069] Step 3: Perform RAG-enhanced multi-level abstraction analysis on the influence argument graph.

[0070] RAG-enhanced multi-level abstraction analysis is performed on the influence argument graph from fine-grained to coarse-grained levels to enhance the contextual understanding of the influence argument graph.

[0071] Step 3.1: Define four levels of abstraction, ranging from fine-grained to coarse-grained generalization:

[0072] (1) (Fine-grained level): The most specific event details, such as the specific search and rescue actions of a single rescue team, the precise time and location, the specific number of casualties and the extent of building damage, and the short-term impact on local traffic.

[0073] (2) (Medium-fine granular level): For The summary includes a series of related rescue operations, coordinated search and rescue within specific blocks, and the medium-term comprehensive impact on the balance of rescue resources in local areas.

[0074] (3) (Medium-coarse granular level): Further abstraction, such as the response deployment of a city's emergency command center, regional evacuation and resettlement goals, and the medium- and long-term profound impact on the overall disaster relief situation.

[0075] (4) (Coarse-grained level): The most macro-level classification, such as the adjustment of national or international organizations' disaster response strategies, global humanitarian assistance goals, and the long-term and profound impact on international disaster management systems and cooperative relationships.

[0076] Step 3.2: For each level of abstraction, retrieve specific contextual information that matches the features of that level of abstraction using the RAG system, including the following steps:

[0077] (1) Submit the description (in text form) of the event to be analyzed to the RAG system as a query.

[0078] (2) The RAG system retrieval tool searches the knowledge base for documents or historical cases that are semantically similar to the event description, or those containing definitions of different abstract levels. - Text with ) characteristics.

[0079] (3) After receiving this retrieved information, the LLM uses its semantic understanding of external knowledge to help determine which abstraction level the event is best suited for. This process introduces the LLM's semantic understanding of a large amount of external knowledge. For example, even if some quantitative indicators of an event are unclear, similar cases retrieved by RAG or clear hierarchical definitions can help the LLM make a more reasonable judgment. Regarding abstraction levels... Search criteria The input is the nodes in the argument graph. and its event description Determine the nodes using the following formula Does it belong to the abstract level? :

[0080]

[0081] Step 3.2: For each level of abstraction, using the Generative Large Language Model (LLM) and the specific contextual information retrieved in Step 3.2, identify and generate an influence transmission pattern representing the characteristics of that level of abstraction, including key elements such as RAG contextual query examples, retrieval behavior, and generator (LLM) behavior. For example:

[0082] (1) Targeting (Fine-grained hierarchical) analysis:

[0083] Example of a RAG query: "Analyze event [X] in..." The study examines hierarchical impact transmission patterns, focusing on aspects such as 'direct damage to critical infrastructure,' and retrieves relevant evidence.

[0084] Retrieval behavior: Retrieves precise details related to event [X] from the knowledge base, such as the type of explosive, the exact location of the explosion, the specific number of casualties, the list of damaged infrastructure, the on-site reports of the first responders, and the specific short-term psychological impact on the surrounding community.

[0085] Generator (LLM) behavior: Based on these highly specific micro-data, it comprehensively describes a very localized, immediate chain of impacts. For example, "An improvised explosive device detonates at [time] at [metro station], causing [X] civilian casualties, directly leading to the disruption of [metro line Y], and triggering [emergency evacuation E of the surrounding area] in the following hours."

[0086] (2) Regarding (Medium-to-fine granular level) analysis:

[0087] Example of a RAG query: "Identify a series of emergency response actions that occurred in [a city] within [24 hours after the typhoon passed]." Hierarchical influence pattern.

[0088] Retrieval behavior: Retrieves multiple items related to this series of response actions. The report includes tiered disaster reports (such as power outages and flooding), the common goals of these response actions (such as restoring basic lifelines), cross-departmental collaborative emergency plans, and a comprehensive assessment report on the pressure on rescue resources in the region.

[0089] Generator (LLM) behavior: combining multiple related The hierarchical influences are aggregated to form The hierarchical model description reveals the synergistic effects and cumulative impacts among these actions. For example, "Within 24 hours of the typhoon's passage, the city's emergency command center successfully restored basic power supply to over 50% of the most severely affected districts and controlled the spread of secondary disaster K by implementing a series of coordinated responses, including emergency drainage, power restoration, and inspection of dilapidated buildings, thus having a medium-term impact on stabilizing the situation in the disaster area."

[0090] (3) Targeting Analysis at the (medium-to-coarse grain level):

[0091] Example of a RAG query: "Analyze the impact of [a city's long-term lockdown A due to a major epidemic] on [regional economy and supply chain]". Hierarchical influence.

[0092] Searcher behavior: Searching for official announcements on epidemic prevention and control, implementation plans involving multiple municipal departments, long-term public health policy intent analysis, and assessment reports on the impact on the entire region's logistics and industries, etc.

[0093] Generator (LLM) behavior: describes a broader, longer-lasting pattern of influence propagation that may connect multiple [factors / initiatives]. The hierarchical model forms a broader strategic picture. For example, "The long-term lockdown A implemented by city [M] aims to achieve the strategic goal of zeroing the virus N. By establishing strict social control P throughout the region, it has significantly changed the economic activity pattern of the area and put long-term pressure on the supply chain stability of surrounding cities Q, forming a regional economic impact pattern that lasts for months."

[0094] (4) Targeting (Coarse-grained level) analysis:

[0095] RAG query example: "Constructing [a comprehensive revision of national building seismic resistance standards implemented after a major earthquake P] and its impact on [the safety resilience of the national construction industry and cities Q] " Hierarchical influence transmission model.

[0096] Searcher behavior: Retrieves relevant national ministry regulations and documents, disaster prevention and mitigation planning outlines, analyses and comments from academicians and experts of the Chinese Academy of Engineering, urban safety plans covering the next few decades, and macro-level research on how major disasters can drive social progress.

[0097] Generator (LLM) behavior: elucidates the most macroscopic, far-reaching, and enduring strategic-level impact transmission patterns, potentially involving the reshaping of international rules and adjustments to the global distribution of power. For example, "the implementation of a comprehensive revision of building seismic standards (P), through the mandatory increase in the seismic intensity of new buildings, profoundly impacts the structure of national urban safety resilience and the evolution of the disaster prevention and mitigation paradigm (U) on a scale of decades. Its transmission path is complex and involves multiple levels, ultimately leading to a qualitative leap (V) in the entire nation's ability to cope with earthquake disasters."

[0098] Taking the argument graph in the embodiment as an example, for event nodes (E1) and The multi-level abstraction analysis of (E8) is as follows:

[0099] (1) E1, Port X in Country A suffered a large-scale cyberattack, resulting in operational paralysis.

[0100] RAG Auxiliary Classification: A system query for "Which abstraction level does event E1 belong to?" might yield documents on topics such as "The impact of cyberattacks on critical infrastructure" and "The economic consequences of port paralysis." After LLM analysis, it may be initially classified as... (The direct physical and economic impacts of a specific event) and (Short-term disruptions to regional trade) level.

[0101] Hierarchical Analysis: RAG retrieves specific technical details about E1 (such as attack methods, duration, affected system modules), estimates of direct economic losses caused by the paralysis of Port X (such as daily value of stranded cargo), and immediate impact on employment in the surrounding area. LLM Generation: "E1, as a precise cyberattack against critical infrastructure, caused the complete shutdown of Port X within 24 hours of its occurrence, resulting in approximately Z million US dollars in direct economic losses per day and temporarily displacing W thousand dockworkers."

[0102] Hierarchical Analysis: RAG search results include an analysis of the importance of Port X in the trade network of Country A and the region, a report on Country B's trade dependence on Port X, and a mid-term study (weeks 1-4) on the impact of similar port disruption events on the supply chain. LLM generation: "The occurrence of E1, due to the key hub status of Port X, rapidly evolved into a significant impact on the regional supply chain." The tiered shock had a significant negative impact, particularly on the economy of Country B, which is highly dependent on this port, for several weeks, manifesting as increased import costs and export delays.

[0103] (2) E8: An international forum proposed adding mandatory cybersecurity clauses to trade agreements.

[0104] RAG Auxiliary Classification: Searching for RAG might yield literature on topics such as "international rule-making," "new issues in global governance," and "the evolution of trade policy." LLM might categorize it as... (Major adjustments to rules in specific areas) and even (A harbinger of a potential paradigm shift in global trade governance) hierarchy.

[0105] / Hierarchical Analysis: RAG retrieves historical cases of major international treaty revisions, international legal discussions on issues such as cyber sovereignty and cross-border data flows, analyses of major countries' positions on the E8 proposal, and predictions of the proposal's potential impact on the global digital trade landscape in the coming years. LLM Generation: "The emergence of the E8 proposal marks the escalation of cybersecurity issues from the technical level to the formulation of international trade rules." / At the strategic level, if widely adopted, this proposal could reshape the compliance framework and entry barriers for global digital trade within the next 5-10 years, profoundly impacting the international trade order.

[0106] Step 4: Perform RAG-enhanced multi-dimensional abstraction analysis on the influence argument diagram.

[0107] Step 4.1: Identify that the nodes in the influence argument graph contain time and location dimension attributes, and define the multiple categories contained in the time and location dimensions as follows;

[0108] (1) Time dimension:

[0109] (Short-term): Duration of less than one week and its short-term impact.

[0110] (Mid-term): Duration from one week to one month, and its mid-term impact on the balance of power, deployment of operations, etc.

[0111] (Long-term): The duration exceeds one month, and its long-term impact on strategic situation, resource allocation, etc.

[0112] (2) Location dimension:

[0113] (Coastal areas): Events involving geographical features such as coastlines and ports and their impact on shipping, maritime defense, etc.

[0114] (Mountainous Areas): Events involving geographical features such as mountains and ranges and their impact on mountain rescue techniques and the access of large rescue equipment.

[0115] (Plains): Events involving geographical features such as plains and grasslands and their impact on large-scale evacuations, the establishment of shelters, transportation lines, etc.

[0116] (Urban Areas): Events involving geographical features such as cities and towns and their impact on infrastructure, social order, urban defense, etc.

[0117] Step 4.2: For the abstract dimension category to which the event node belongs, use the RAG system to retrieve contextual information related to that dimension category, including the general influence mechanism or historical cases of the influence transmission of that dimension category. For the time dimension... Search criteria The input is the nodes in the argument graph. and its event description and time information The output is a boolean value, representing the node. Does it belong to the time dimension? The event, the formula is:

[0118]

[0119] For location dimension Search criteria The input is the nodes in the argument graph. and its event description and location information The output is a boolean value, representing the node. Does it belong to the location dimension? The event, the formula is:

[0120] At this point, the core value of the RAG system lies in its ability to not only determine which dimension category an event node belongs to (e.g., "this is a..."). (Short-term events), and more importantly, it can retrieve general impact mechanisms for this dimension category from the knowledge base, as well as deeper knowledge such as common impact patterns of similar dimension events in history. This step includes RAG context query examples and retriever behavior, as shown below:

[0121] (1) Time dimension ( - RAG system retrieval:

[0122] RAG query example (for) ): "Analyze event [X] (identified as The impact transmission of short-term events. Retrieve knowledge bases regarding typical impact patterns (e.g., impact, duration, spread, etc.) of short-term events (within one week), as well as historical cases of short-term events similar in nature to [X] and their impacts.

[0123] Retrieval behavior:

[0124] for (Short-term): Search for information on emergency response mechanisms, cases of rapid information dissemination in the short term (such as panic selling on social media), data on short-term market fluctuations of emergency supplies, and studies on the immediate impact on public sentiment.

[0125] for (Medium-term): Search for reports on the effects of policy adjustments over the medium term (weeks to one month), analyses of the formation of community functions in temporary resettlement sites in disaster areas within one month, and the response of medium-term economic indicators (such as regional unemployment rates) to specific events.

[0126] for (Long-term): Search for long-term social impact assessments (from months to years) of major post-disaster reconstruction projects (such as rebuilding schools and hospitals), longitudinal studies on the impact of long-term policies (such as revising emergency plans) on social structures, and analyses of the impact of long-term factors such as climate change on disaster patterns.

[0127] (2) Location dimension ( - RAG system retrieval:

[0128] RAG query example (for) Mountainous area): "The event [Y] occurred in the mountainous area ( The search will examine the typical constraints and advantages that mountainous geographical environments impose on disaster relief, material transportation, information communication, and public self-rescue, as well as historical case studies of the transmission of similar [Y] events in mountainous areas.

[0129] Retrieval behavior:

[0130] for (Coastal): Search for information on the pivotal role of ports in international trade, their vulnerability to storm surges and tsunamis, the vulnerability of coastal cities to sea-level rise, and historical cases of using ports for large-scale humanitarian relief efforts.

[0131] for (Mountainous Areas): Search for information on special techniques for mountain rescue (such as rope rescue), the limitations of complex terrain on the passage of large rescue equipment, the challenges of communication signal coverage in mountainous areas, and historical cases of using mountainous areas as refugees for disaster victims or sources of secondary disasters.

[0132] for (Plains): Search for information about the suitability of plains for large-scale evacuation and the establishment of shelters, the importance of transportation lines, historical large-scale floods in plains, and the strategic characteristics of plains as easy to defend but also vulnerable to disasters.

[0133] for (Cities): Search for information on the complexities of urban search and rescue, the vulnerability of critical infrastructure (water, electricity, communications), the impact of urban population density on public opinion dissemination, historical cases of urban terrorist attacks or major accidents, and the importance of cities as political, economic, and cultural centers.

[0134] Step 4.3: Using the aforementioned Generative Large Language Model (LLM), combined with the contextual information retrieved in Step 4.2, analyze and explain how the abstract dimension affects the event propagation pattern, as shown in the following example:

[0135] (1) Time dimension ( - RAG enhancer analysis

[0136] Generator (LLM) behavior: LLM uses this retrieved information to explain why the currently analyzed event [X] is due to its specific temporal attributes ( / / This results in certain characteristics of influence transmission. For example, for a... For events, LLM may combine the general pattern of "short-term events having a large impact but a rapid decay" found in the search, as well as similar historical cases, to analyze the intensity and duration of the impact of [X].

[0137] (2) Location dimension ( - RAG enhancement analysis

[0138] Generator (LLM) behavior: LLM integrates retrieved geographic feature knowledge and relevant cases to analyze how the impact of event [Y] is affected by its specific location attributes ( - Shaped by ) . For example, for what happens In mountainous areas, LLM may point out that the speed of the impact's spread is limited, but it may be more covert, more persistent, and more likely to form small, difficult-to-eradicate centers of influence.

[0139] Taking the argument graph in the embodiment as an example, the multi-abstract dimension analysis of event node E1 is as follows:

[0140] (1) Analysis of the impact of E1 in the time dimension:

[0141] (Short-term): RAG retrieves news flashes, social media responses to the port paralysis, and emergency response statements from Country A's government within a week of E1's occurrence. LLM describes: "Following E1, information spread rapidly through global media ( (Characteristics) The government of Country A activated an emergency response within 48 hours, but failed to prevent panic from spreading among importers in Country B.

[0142] (Medium-term): RAG retrieved shipping data from 2-4 weeks after E1, monthly trade deficit reports from Country B, and official documents from the International Shipping Organization (ISO) regarding the adjustment of risk ratings (E3). LLM analysis: "Medium-term impact of E1 ( This is reflected in Country B's month-long import difficulties and rising costs, as well as the increased risk perception in the international shipping market regarding the region (E3).

[0143] (Long-term): RAG retrieved reports on the progress of trade negotiations between Country B and Country D several months after E1 (E7), think tank reports on global supply chain resilience, and Country A's long-term budget for repairing Port X and upgrading cybersecurity. LLM concluded: "The long-term impact of E1 ( One of the reasons is that it will prompt Country B to seek supply chain diversification (E7), and may trigger long-term global investment in raising cybersecurity standards for critical infrastructure.

[0144] (2) Location dimension analysis E1 (occurred in port X of country A, belonging to Coastal areas):

[0145] RAG query: "Port X as..." How does the geographical location of coastal facilities affect the transmission of E1 incidents?

[0146] Searcher behavior: Search for information about the geographical and strategic location of port X (e.g., whether it is a vital shipping choke point), its share in country A's foreign trade, vulnerability analysis of coastal areas to maritime blockade or attack, and historical cases of similar attacks on key coastal nodes.

[0147] LLM's actions: LLM explains: "E1's ability to generate such a wide international influence is partly due to Port X ( It is a crucial maritime gateway for Country A and even the entire region. Its coastal location makes it an easy target for external attacks; if it is paralyzed, the impact would quickly spread along international trade chains that heavily rely on maritime transport to trading partners such as Country B.

[0148] Step 5: Perform RAG-enhanced variable step size influence path analysis on the influence argument graph.

[0149] A path analysis of direct and indirect effects with variable step sizes is performed, and the RAG system is used to retrieve evidence of causal relationships supporting the connections between nodes in the effect path.

[0150] Step 5.1: Define two types of step sizes for influence path analysis: long step size ( ) and short step size ( Long step length focuses on the macroscopic impact between the starting node and the final node; short step length focuses on the gradual transmission path including intermediate nodes.

[0151] (1) Long step length This only considers the direct influence transmission relationship between the starting node and the ending node, ignoring intermediate nodes. For example, for the transmission chain a→b→c→d→e, Focus only on the overall effect of a→e.

[0152] (2) Short step length Consider the connections from the starting node to the final node via some intermediate nodes. For example, for the propagation chain described above, Possible paths include a→b→e, a→c→e, or a→d→e.

[0153] Step 5.2: Path Discovery and Verification Assistance. Given a starting node and a final node, the RAG system is used to perform a graph search algorithm in the argument graph to initially identify candidate paths. Textual evidence supporting the existence of candidate paths is searched from the knowledge base, and paths conforming to long step sizes are discovered and verified. ) or short step size ( The potential impact path is defined by the step size. Search criteria The input is a debate diagram. Starting node and end node The output is a combination of nodes. and its influence transmission relationship The formula is:

[0154]

[0155] The impact path discovery and verification assistance includes key aspects such as RAG context query examples, retriever behavior, and generator (LLM) behavior, as shown in the following examples:

[0156] RAG query example: Given a starting event (node) ) and a target event (node) When this happens, you can query the RAG system to see if there is a relationship between them. or The potential impact path of the type.

[0157] Retrieval behavior: First, the retrieval can perform graph search algorithms (such as variants of BFS, DFS, or more advanced pathfinding algorithms) on the indexed argument graph to initially identify candidate paths.

[0158] Then, for each candidate path (regardless of) still The retrieval engine searches the knowledge base (text data) for evidence supporting the existence of the path (or the connections between segments in the path). For example, for the path a→b→e, the retrieval engine will look for documents that explicitly state the facts that "a leads to b" and "b leads to e".

[0159] If causal reasoning capabilities are integrated, the retrieval system will pay special attention to texts containing causal indicators (such as "cause", "because", "therefore", "cause"), or use causal knowledge graphs extracted from the text to verify the causal rationality of the path.

[0160] Generator (LLM) behavior: LLM evaluates the rationality and strength of these candidate paths based on retrieved graph structure information and textual evidence.

[0161] for Path analysis, in LLM, determines whether a direct macroscopic link from the starting event to the final event truly exists, or is merely a temporal sequence. High-level analysis reports or expert comments retrieved from RAG may support or refute this direct link.

[0162] for The path, LLM examines the connection strength and logic of each intermediate link to ensure that the path is not far-fetched.

[0163] Step 5.3: Evidence Enhancement for Path Influence Mechanisms. Using the Large Language Model (LLM) and the path evidence retrieved in Step 5.2, assess the significance of the influencing path and elucidate its transmission mechanism, including key aspects such as RAG query examples, retrieval behavior, and generator (LLM) behavior. Examples are as follows:

[0164] RAG query example: Once an impact path is initially confirmed (e.g., through...) The analysis yields the path a→b→e, and RAG can further provide detailed contextual information for each node and each connection segment (a→b, b→e) along this path.

[0165] Retrieval behavior: For nodes a, b, and e, retrieve their detailed background, attributes, and roles in other events. For connections a→b and b→e, retrieve the specific mechanisms explaining how this influence occurred, relevant conditions, the strength and type of the influence, etc.

[0166] Generator (LLM) behavior: LLM uses this rich detail to explain how the influence is specifically transmitted along this path, and what bridging or amplifying / reducing role the intermediate link b plays. This deepens the understanding of the influence path from "knowing it exists" to "understanding why and how it happens."

[0167] Taking the argument diagram in the embodiment as an example, and The analysis of the influence on the transmission path is as follows:

[0168] (1) Long step length analyze: (E1, Cyberattack → E7, Country B Seeks New Alliance with Country D)

[0169] RAG query: "Is there direct evidence or high-level analysis linking the cyberattack on country A's port (E1) to country B's eventual pursuit of closer trade relations with country D (E7)?"

[0170] Searcher behavior: Search for possible think tank report titles such as "How the Port X Incident Reshaped the Regional Trade Landscape", and speeches by officials from country A or country B that, months later, mentioned that E1 was a "catalyst" that led to adjustments in country B's foreign policy.

[0171] LLM Behavior: LLM assesses the strength of this macro-level link. If strong evidence is retrieved, the S1 path is confirmed. For example: "Although E1 and E7 are separated by several months, multiple strategic analyses retrieved by RAG indicate that the supply chain vulnerabilities exposed by E1 were a key driver for Country B to reassess its trade dependence and ultimately proactively engage with Country D (E7), constituting a strategic-level link." The path of influence.

[0172] (2) Short step length analyze: (E1, Cyberattack → E2, Trade Disruption → E6, G7 Meeting → E8, New Provisions Proposal)

[0173] RAG's research into news reports from Country B regarding complaints from domestic importers following E1, and official statements from Country B regarding trade disruptions, confirms that "E1 directly led to E2, and the evidence is strong."

[0174] RAG's analysis examined whether any G7 member states (particularly Country B or its allies) publicly called for G7-level discussion of the matter following E2, and whether the G7 meeting agenda explicitly mentioned trade security issues arising from E1 / E2. LLM's analysis stated: "Country B's economic losses due to E2 gave it a strong incentive to push for discussion on the G7 platform (E6), and RAG's findings showed that Country B's foreign ministers issued statements supporting this connection before the meeting."

[0175] RAG searched the outcome statement or press conference records of the G7 meeting (E6) to see if they mentioned promoting relevant rule-making at an international forum (where E8 took place). LLM confirmed / disproved: "The joint statement of the G7 meeting (E6) clearly stated its commitment to strengthening global supply chain cybersecurity and authorized its representatives to propose specific solutions at [a forum], which is directly related to the emergence of the E8 proposals."

[0176] Step 6: Calculation of the impact strength of RAG contextualization

[0177] The RAG system is used to calculate the influence strength of contextualization, which serves two purposes: first, to dynamically enrich or adjust the representation of the semantic vectors of each node; and second, to provide evidence to support the determination of the weights of logical relationships.

[0178] For the combination of path nodes Node sequence in Influence on transmission relationship Calculate the influence strength between nodes To represent, the formula is:

[0179]

[0180] in, and It is a node and The components, represented by a certain semantic vector, are converted into semantic vectors through natural language processing to transform node information. It is a logical relation weight, determined based on the causal relationship and logical connection between events. For example, if event Directly caused the incident , Take the larger value, and vice versa.

[0181] Taking the argument diagram in the embodiment as an example, (E1, cyberattack → E2, trade disruption) The impact transmission path analysis of (E5, Country C denies Country A's accusations → E9, Tensions arise between Country A and Country C) is as follows:

[0182] (1) Calculation :

[0183] Semantic Vectors: RAG retrieves extensive detailed descriptions of E1 (the port was paralyzed by a cyberattack) and E2 (trade between country B was disrupted), generating context-aware embedding vectors for this specific connection.

[0184] Logical weighting: RAG retrieved numerous news reports and official data confirming that E1 is the direct and primary cause of E2. Following LLM analysis, it is determined that... The logical weight is assigned a high value close to 1.

[0185] Calculated The value will be very high, reflecting a strong direct impact.

[0186] (2) Calculation :

[0187] Semantic Vector: RAG retrieves the full text of Country C's denial statement (E5) as well as subsequent statements of mutual accusations between the spokespersons of the foreign ministries of A and C, the official announcement of the recall of the ambassadors (E9), etc.

[0188] Logical weighting: Information retrieved by RAG indicates that E5 (denial) is a significant factor leading to E9 (escalation of tension), but other factors may also exist (such as the accusation itself in E4). LLM will conduct a comprehensive evaluation and assign weights accordingly. The logical weight is a medium to high value.

[0189] Step 7: Generate the influence transmission model

[0190] LLM synthesizes the multi-perspective analysis results enhanced by RAG to generate a comprehensive impact transmission model of the "Cyberattack on Port A in Country A". This model will not only be a graph, but also a rich text report, including:

[0191] (1) Core transmission path: For example, showing the process from the point of view with pictures and text. Starting from (E1), through various intermediate events, it ultimately leads to the following... (E7) (Policy Shift in Country B) (E8) (New Proposal for International Rules) (E9) (Deterioration of diplomatic relations) and other key long-term consequences ( ).

[0192] (2) Interpretation of the influence of multiple levels of abstraction: Explaining E1 at different levels of abstraction ( - The impact of )

[0193] (3) Interpretation of multi-abstract-dimensional features: Analysis time ( - ) and location ( How multiple abstract dimensions of factors shape the way and scope of influence spreads.

[0194] (4) Quantitative assessment of key nodes and connections: Provide strength values ​​of connections that significantly affect the connection. And explain the basis for it (evidence from RAG).

[0195] Finally, based on the impact transmission model, the study summarized the patterns and, through LLM, extracted the universally significant impact transmission patterns observed in this event. For example, "Critical infrastructure (especially coastal hubs) subjected to cyberattacks can, in the short term (…)" Within a short period, the economic impact can be rapidly transmitted to trading partners through the supply chain, and may extend into the medium to long term. - This has triggered affected countries to seek geoeconomic adjustments and promote changes in relevant international rules. - ).

Claims

1. A method for discovering influence transmission patterns based on multi-view and retrieval-enhanced graphs, characterized in that, Includes the following steps: Step 1: Construct an impact argument graph based on event data. The nodes of the impact argument graph represent events or impacts, and the edges represent the transmission of impacts. Step 2: Establish a RAG system, which includes a knowledge base storing structured and unstructured information related to the event data, a retrieval module, and a generative large language model; Step 3: Based on the RAG system, perform RAG-enhanced multi-level abstraction analysis on the influence argument diagram; Step 4: Based on the RAG system, perform RAG-enhanced multi-dimensional abstraction analysis on the influence argument diagram; Step 5: Based on the RAG system, perform RAG-enhanced variable step size influence path analysis on the influence argument graph; Step 6: Based on the RAG system, perform RAG-contextualized influence strength calculation on the influence argument graph; Step 7: Based on the impact patterns and impact paths identified in Steps 3-6, generate a comprehensive model of the transmission of the impact of major events.

2. The method for discovering influence transmission analysis patterns according to claim 1, characterized in that, In step 3, multiple levels of abstraction are defined, ranging from fine-grained to coarse-grained generalization. RAG-enhanced multi-level abstraction analysis is performed on the impact argument graph at different granularities to enhance the contextual understanding of major emergency events and their impacts in the impact argument graph.

3. The method for discovering influence transmission analysis patterns according to claim 2, characterized in that, In step 3, regarding the level of abstraction The retrieval module of the RAG system is used to retrieve specific contextual information of the argument graph, including: targeting nodes in the argument graph. Retrieval and Abstraction Level Feature matching and supports nodes Divided into abstract levels Specific contextual information; for the level of abstraction Search criteria The input is the nodes in the argument graph. and nodes Corresponding event description The output is a boolean value, representing the node. Does it belong to the abstract level? The formula for judging the event is: 。 4. The method for discovering influence transmission analysis patterns according to claim 3, characterized in that, In step 3, regarding the level of abstraction By utilizing the generative large language model of the RAG system and the specific contextual information, the abstraction levels of representation are identified and generated. The influence transmission pattern of characteristics; once an event node is categorized into a specific level of abstraction, the generative large language model analyzes the level of abstraction. The impact patterns are analyzed and descriptions are generated; the analysis at each level of abstraction includes RAG query examples, retriever behavior, and generator behavior.

5. The method for discovering influence transmission analysis patterns according to claim 1, characterized in that, In step 4, the RAG system is used to identify multiple predefined abstract dimension attributes that affect the nodes of the argument graph. The abstract dimensions include at least time and location dimensions, and each abstract dimension contains multiple predefined categories. For each specific abstract dimension category to which an event node belongs, the retrieval module of the RAG system is used to retrieve contextual information related to that specific abstract dimension category, including the general influence mechanism or historical cases of the influence transmission of that dimension category.

6. The method for discovering influence transmission analysis patterns according to claim 5, characterized in that, In step 4, the generative large language model is used, combined with the contextual information related to the specific abstract dimension category, to analyze and explain how the specific abstract dimension affects the event transmission pattern. The analysis for multiple abstract dimension categories includes RAG query examples, retrieval behavior, and generator behavior.

7. The method for discovering influence transmission analysis patterns according to claim 1, characterized in that, In step 5, define at least a long step size. and short step length Step size for analysis of multiple influence paths ( The system performs direct and indirect impact path analysis with variable step size. Using the RAG system, it retrieves evidence of causal relationships supporting the connections between nodes in the impact path. The long step size focuses on the direct and macroscopic impact transmission relationship between the starting node and the final node, ignoring intermediate nodes. The short step size focuses on the asymptotic transmission path that includes intermediate nodes, considering the connection from the starting node to the final node through some intermediate nodes.

8. The method for discovering influence transmission analysis patterns according to claim 7, characterized in that, In step 5, for a given starting and ending node, the RAG system is used to assist in the discovery and verification of influence paths. Candidate paths are identified in the argument graph using a graph search algorithm, and the retrieval module of the RAG system is used to search the knowledge base for textual evidence supporting the candidate paths, discovering and verifying that they conform to the analysis step size of each influence path. Defined potential impact paths; The influence path discovery and verification assistance includes RAG query examples, retriever behavior, and generator behavior, for step size... Search criteria The input is a debate diagram. Starting node and end node The output is the combination of path nodes. and the influence transmission relationship The calculation expression is: 。 9. The method for discovering influence transmission analysis patterns according to claim 8, characterized in that, in step 5, the generative large language model is used in conjunction with the search for textual evidence supporting candidate paths from the knowledge base to evaluate the transmission mechanism of path influence; the evidence enhancement of the transmission mechanism of path influence includes RAG query examples, retrieval behavior, and generator behavior.

10. The method for discovering influence transmission analysis patterns according to claim 8, characterized in that, in step 6, the RAG system is used to calculate the contextualized influence intensity for combinations of influence path nodes. Node sequence in Influence on transmission relationship Calculate the influence strength between nodes To indicate, the The calculation formula is as follows: ， in, and It is a node and The components, represented by a certain semantic vector, are converted into semantic vectors through natural language processing to transform node information. It is the logical relation weight, which is determined based on the causal relationship and logical connection between events.

Images